V-JEPA 2.1 ViT-L/16 (384) — HF port

HuggingFace-format weights for V-JEPA 2.1, ViT-L/16 @ 384px (distilled from ViT-G), for dense spatio-temporal feature extraction.

• patch 16, tubelet 2 → a 64-frame 384² clip yields a 32 × 24 × 24 token grid (D=1024).
• Converted from Meta's official checkpoint vjepa2_1_vitl_dist_vitG_384.pt using the HF port at github.com/Dev-Jahn/vjepa2-hf.

Usage

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "Dev-Jahn/vjepa2.1-vitl-fpc64-384", trust_remote_code=True,
).eval().cuda()

# channels-first (B, C, T, H, W); H=W=384, ImageNet-normalized
pixel_values_videos = torch.randn(1, 3, 32, 384, 384, device="cuda")
with torch.no_grad():
    feats = model(pixel_values_videos, skip_predictor=True).last_hidden_state  # (B, N, 1024)

# dense grid: N = (T/2)*(384/16)*(384/16)
grid = feats.reshape(feats.shape[0], 16, 24, 24, -1)  # for T=32

Attribution & license

Derived from Meta's V-JEPA 2.1 (MIT/Apache-2.0). Port code and this conversion are Apache-2.0. See the original facebookresearch/vjepa2.